On Friday evening, 4:00, Anthropic dropped a bomb. They said you can no longer use thirdparty harnesses, aka OpenClaw, with your Claude subscription. And yes, I'm on vacation, but I just had to talk about this because it is so nuts and there's so much to discuss here. All right, let's look at the email first. Here it is. So, as of April 4th at 12:00 p.m., we're enforcing the new policy we shared previously. So, here's the thing. They sent the email the day before, less than 24 hours before, saying that if you're using a thirdparty harness with your Claude subscription, it is now definitively against their terms of service. But it actually is more confusing. But I'm going to get to that in a minute. And they specifically called out third party harnesses, including OpenClaw. They specifically said OpenClaw. They are shunning the OpenClaw community, and it will no longer draw from your claw subscription usage limits. to keep it going, you'll need to turn on extra usage. Basically, you're going to be paying a lot more. And maybe it makes sense for their

business model because cursor actually put out a report that says a $200 cloud subscription is actually about $2,000 worth of credits from Anthropic. And so, this is all VCback subsidization of tokens. And they tried to soften the blow a little bit. If this change doesn't work for how you use Claude, you can cancel your subscription and get an automatic refund. So, they will give you a full refund, not just your unused credits. Boris Churnney, the head of Claude Code, said the same thing and then continues with, "We've been working hard to meet the increase in demand for Claude and our subscriptions weren't built for the usage patterns of these third-party tools. Capacity is a resource we manage thoughtfully, and we are prioritizing our customers using our products and API. You get a onetime credit equal to your monthly plan and if you need more, you can buy discounted usage. Okay, so look, Anthropic is going through something right now. Two things simultaneously. Their usage is going crazy and I'm going to show you that in a minute. But also, they don't have

enough capacity and Anthropic has been in such a GPU crunch, they've actually used both the stick and the carrot approach to try to lower usage or at least balance usage a little bit better. So, here is the first carrot. A small thank you to everyone using Cloud. We're doubling usage outside our peak hours for the next 2 weeks, and I believe it's still going on right now. So, 2x usage on weekdays outside of 5 to 11 Pacific and 2x usage all day on weekends, automatic, nothing to enable. The interesting thing, keep this window in mind, 5 to 11 a.m. Then the big one, the stick. Here's Tariq. To manage growing demand for Cloud, we're adjusting our 5-hour session limit for free pro max subs during peak hours. Your limits remained unchanged. Here it is. During weekdays, 5 to 11 Pacific, you'll move through your 5-hour session limits faster than ever. So, they've landed a lot of efficiency wins, but this will affect 7% of users, aka Open Claw users,

most likely, agentic users most likely. But this wasn't even enough. Even after this, they still started banning OpenClaw as of a couple days ago. But that's not it. People have been seeing their quotas just explode overnight, basically barely using their Cloud subscription and then running out within a day or two of the week resetting. So, a lot of people are having problems with this. Okay, so now back to this email. Basically, you have less than 20 hours. It's about 19 hours since receiving this email to change your open claw primary model. You can no longer use clawed models there. At least I don't think so. Again, more on that later. But they did specifically call I got this rooster out there that won't stop making noise during this video. He's just as frustrated as I am. But here's the thing. It's actually very easy to swap out your models. There is effectively no switching cost to Frontier models. Zero. In fact, let me show you a tweet I put

out and let me show you who responded and agreed. So, here's my post. There is literally zero switching cost for models. In fact, right when I received that email, here's what I did. Switch anywhere we're using cloud models to use GPT 5.4 thinking instead through the codeex API. I meant codeex ooth instead. But, of course, OpenCloud knew what I meant and did it the right way anyway. Make sure to switch the prompt files for the ones optimized for GPT. I'm on it switching. Okay, literally like 3 minutes later, all of my clawed models in OpenCloud were swapped out for GPT 5.4. It was that easy. I didn't have to do anything. But let me point out one thing in particular. Make sure to switch the prompt files for the ones optimized for GPT. I've talked about this before, but I just want to touch on it really quickly. I have multiple variations of every single prompt file in my open clause system. That makes it really easy to swap models around, but more importantly, I optimize the prompts for the models. So, for example, a prompt for Opus 4.6 looks very different than a

prompt doing the same exact thing for GPT 5.4. So, just keep that in mind. Little tip. And Mr. Jack Dorsey, creator of Twitter, agreed with me. Yes. So, he agrees. There's just literally zero switching cost. So even though Enthropic gave us very little time to swap out our models in OpenClaw, it was just dead simple. There really isn't that much to it, especially if you already have the prompt file optimizations in place. If you didn't, takes a few extra minutes at most. And by the way, if you want to run some of this on open source models, Digital Ocean is a great option and the sponsor of today's video. Building and scaling AI in production is harder than it should be. If you watch this channel, you know that better than anybody. You're either going to a hyperscaler that is overly complex or you're building on raw GPUs which have no software and you're doing everything yourself. Meanwhile, specialized inference rappers are merely software layers renting GPUs from Neocloud middlemen. But there is a better way.

Digital Oceans, a Gentic inference cloud, is purpose-built for production AI. I've actually been using Digital Ocean for a long time in my previous businesses as well. So super excited to tell you about them today. With Digital Ocean's Agentic inference cloud, you get the best of both worlds. You get reliability and simplicity. And this is inference optimized infrastructure. If you're an AI developer, you get everything you need to deploy AI models, scale to millions of requests, and operate reliably in production without the complexity or cost of a hyperscaler, and without having to write all the software yourself, like going to a bare metal GPU. AI companies running at high volume are already experiencing gains in throughput, latency, and cost efficiency. So, if you're building and scaling AI in production, check out Digital Ocean's Aentic inference cloud today. Tell them I sent you. Link down below. And to make things even more confusing because Anthropic cannot seem to add clarity to their policies

whatsoever, Shamir Webbby, sorry if I'm pronouncing your name right, said, "What about agents SDK?" That's what I've been using to use my cloud subscription with OpenClaw. And Boris Churnney said, "No changes to agents SDK at this time. Working on improving clarity more." This was a few days ago. We still don't have clarity. We still don't know. So, is it okay to use the agents SDK within the OpenClaw ecosystem? I don't know. At this point, it's not even worth taking the risk. It's not even worth using it because at the same time that Enthropic is trying to cut your quota, at the same time Anthropic is having this GPU crunch, OpenAI is resetting their quotas all the time. They have very liberal quota policies. I don't think I've ever come up against a codeex quota. So, just keep that in mind. But GPT 5.4 doesn't have the best personality, especially with an OpenClaw. But the OpenClaw team has been working on that. Let me show you what it looks like to be in a GPU crunch at all times. This is Claude's

status page. Look at all these reds. Their total uptime is 98.77% for claw.ai. I know that sounds really high, but if you don't have 99 plus, that is effectively unusable. The Claude API dropped below 99%. Claude Code still sitting above 99%. Platform still sitting above 99%. And if you thought all of this means they aren't growing that quickly, you would be so wrong. Look at this. Anthropic boasts revenue run rate of 30 billion, finally surpassing OpenAI's revenue run rate, at least what we know publicly. And they've actually landed a huge deal with Google to basically use their TPUs. So at the end of 2025, Anthropic's revenue run rate was 9 billion and now it's 30 billion. They are seeing absolutely vertical growth on their revenue and that's why they're in such a GPU crunch. They're doing everything they can to keep up with demand and they can't. So if you thought AI was a fad,

if you thought AI was hype, just look at these graphs. Obviously, the coding use case alone is worth 30 billion ARR because that is what Anthropic is focused on almost entirely. This is what their run rate looks like. Look at this absolutely vertical line right here. Just so crazy. And so, let me talk about this for a second. If you are building anything with agents, you need a multimodel strategy. And I'm not just talking about multi-frontier model. You should be incorporating local models as well. Now obviously local models can't do the kind of highle orchestration and planning as well as a frontier model. They can't really do coding as well as a frontier model. But that's okay. You can actually offload large parts of your agent stack to open source models. Classification, data extraction, summarization. There are a number of different use cases, especially with Gemma 4 now coming out that's really good. We have Quen 3.5 that's really good. And I have a video coming soon about how to do that. Subscribe if you aren't already to make sure you get

that. And apparently there is now a classifier that is blocking third-party harnesses and what Peter says first-party harnesses. Look at this. Anthropic now blocks first party harness use too. Okay, Claude. If you append the system message, a personal assistant running inside OpenClaw is Claude here. Boom. Third party apps now draw from your extra usage, not your plan limits. So, Anthropic is literally banning prompts. literally banning props. This is ridiculous at this point. However, Boris Churnney jumped in. This is not intentional. Likely an overactive abuse classifier looking and working on clarifying the policy going forward. They still haven't clarified it. It's actually been super unclear for weeks and weeks now. What is allowed? What is not allowed? Is the agents SDK allowed or not? Is the agents SDK with extra usage allowed or not? Why is it so complicated? Why can't they just clarify it? Just give us a policy and change it in the future if you need to, but give us something now. The lack of clarity is

beyond frustrating. So look, this is frustrating. Yes, it is increasingly difficult to depend on anthropics models because of their changing policies, their unclear policies. But at the same time, we have Open AI who quote unquote acquired Peter Steinberger who says come use OpenClaw with us. Of course they do. Of course they want to. And the reputation gained from Peter Seyberger joining OpenAI and Anthropic shunning OpenClaw users is palpable. And OpenAI resetting quotas all the time and OpenClaw being made to work well with GPT 5.4's personality or at least making GPT 5.4's personality usable. So we're all headed to GPT 5.4. It seems open AAI is doing the right thing. anthropic. They are being frustrating to say the least, but they're doing what's best for their business. They're generating a ton of revenue. They have hyperfocus on reaching AGI first. So, all of this

OpenClaw stuff, it doesn't matter to them. That's the thing. It doesn't matter to them. And here's Peter Steinberger's tweet saying, "One of the biggest upgrades in the latest version of OpenClaw is the fact that it makes GPT 5.4's personality actually good. It actually feels right. It feels claw." So, anyone who want to help me test the GPT 5.4 upgrades and someone says, "Okay, works well for me. Getting the old clawed vibes. Thanks a lot. It took a hammer, but I managed to beat emotions into GPT." So, very cool. Very good work, Peter. But again, multimodel. Don't depend on one model. Get the most out of every model. That is the way. If you enjoyed this video, please consider giving a like and subscribe. and I'll see you in the next